Gemma 4 ran an inference in orbit, but some don't accept the 'first space LLM' claim

Sundar Pichai announced on X the 'hello world' milestone for Project Suncatcher. A few weeks after launching the prototype satellite, this time they sent the command to orbit, started the TPU, ran the prompt, and received the first complete response on the Earth side: a haiku.

Philipp Schmid reposted the message and said Gemma 4 had become the first LLM to run in space. Gemma is Google's open-source model family, and TPU is Google's self-developed AI chip. The repost pushed the discussion from an engineering demonstration to the question of what counts as 'first'.

Sundar's original post was written in a restrained way: send the command, start the TPU, run the prompt, receive the complete response. The focus was on the completeness of the process; the orbital side did not just send back a heartbeat, but completed one model inference. As for whether the model was Gemma 4 and how this 'first' should be counted, the comment section had no unified answer.

Ahmad Al-Dahle said in a reply that Llama 3.2 was the first named model to run in space, in April 2025. Philipp replied asking whether he had an article, saying that he had actually missed it. 3DTOPO laid out a longer timeline: in 2024, a lightweight LLM ran on HPE's Spaceborne Computer-2 inside the International Space Station; in 2025, 'Space Llama' ran on the same hardware; in 2025, nano-GPT trained and performed inference on an H100 GPU; in 2026, Satlyt quantized the Gemma 3 1B model for inference.

These statements split 'first' into different definitions: is it the first LLM to run in orbit, the first named model, or the first time an end-to-end command plus inference plus return has been completed? What Project Suncatcher can confirm this time is that the link worked.

In the comments, Raghav Mehta asked what would happen if radiation flipped one bit in the weights; Henkjan asked what space constraints actually tested that ground simulations could not catch; pranjulya joked that choosing a haiku for the first output was very fitting—short enough to survive the downlink, and very hard to fact-check. It was a joke, but it also touched on the real limits of space inference: downlink bandwidth, result validation, and hardware reliability all have to be counted in.

Cassidy Waters's reply came with a robot-shaped video, which really fit the vibe of this round of discussion: the cloud has already been put in the cloud; next, what matters is how long it can run stably in orbit.

Original links: Philipp Schmid's tweet, Sundar Pichai's original tweet.